System Architectures Using Network Attached Peripherals Rodney Van Meter USC/Information Sciences Institute rdv@isi.edu http://www.isi.edu/netstation/ USC Integrated Media Systems Center Student Council Seminar October 30, 1997 1
Talk Outline • Introduction • Network Technologies • NAPs in Multimedia • NAPs in Mass Storage • Operating System Support • Conclusion 2
What is a Network Attached Peripheral? Any computer peripheral attached directly to some form of network, rather than a bus. • HiPPI frame buffers • Fibre Channel disk drives • ATM cameras 3
Characteristics of Network-Attached Peripherals (NAPs) • Scalable physical interconnect (# of nodes, distance, etc.) • No physically defined owner • Interconnect shared w/ general-purpose traffic • Higher latency • Delivery subject to usual network problems (packet loss, out-of-order delivery, fragmentation, etc.) • Support for 3rd party transfer (direct device-to-device communication) Present in varying degrees in different systems. 4
Problems Faced with NAPs Closed, bus-centric architecture allows simplifying assumptions about resource identification, security and sharing. • Set of resources not constrained by architecture • Network issues of scale & heterogeneity • Control of devices not limited to bus master • Non-dedicated network • Security now paramount 5
What are NAPs Good for? • Better scaling (distance, # nodes, aggregate bandwidth) • Simpler cabling • Direct device-to-device communication • Direct device-to-client comm. reduces server load 6
Talk Outline • Introduction • Network Technologies • NAPs in Multimedia • NAPs in Mass Storage • Operating System Support • Conclusion 7
Network Technologies for NAPs All seven layers in ISO model open to debate • Application • Presentation • Session • Transport • Network • Link • Physical 8
Proposed & In-Use Networks • HiPPI 800 • HiPPI 6400 • Fibre Channel fabrics • Fibre Channel Arbitrated Loop • FireWire (1394) • Gigabit ethernet • ATM • Serial Storage Architecture (SSA) • Myrinet • various others 9
High Performance Parallel Interface (HiPPI) • Goals: simple & fast (800 Mbps), supercomputing • Switched or routed • Parallel copper or serial fiber • Phy, link layers • IPI-3 or TCP • Weaknesses: limited scalability 10
Fibre Channel • Goals: fast, scalable, distance (ambitious) • Serial copper coax or fiber • 800 Mbps • Switched fabric or arbitrated loop • Phy, link, net, transport layers • SCSI commands over custom transport • Front runner for “winner” • Weaknesses: expense, complexity; scalability and loop/fabric interoperability unproven (low pkt loss rate, in-order delivery assumptions may not hold) • http://www.fibrechannel.org/ 11
FireWire 1394 • Goals: simplicity, low cost, desktop environment • Custom copper cables • 100, 200, 400 Mbps • Arbitrary physical topology, but shared/broadcast medium • Phy, link, net, transport layers • Very bus-like • Weaknesses: shared low bandwidth; nothing scales • http://www.firewire.org/ 12
Gigabit Ethernet • Goals: interoperability w/ ethernet switches, similar programming model • Tweaked Fibre Channel physical • 1 Gbps • Phy, link layers • Likely popular for GP traffic, can it translate to storage? • Weaknesses: small packet size, expense, undefined storage profile 13
Networking Problems for NAPs HiPPI-6400 ❁ gigabit Ethernet ❈ Myrinet ❊ FC-AL as I/O Nets Get Larger and More Complex: 1394 ❅ HiPPI-800 ❃ ATM ❄ SSA ❉ Fibre Channel • Media Bridging (Routing, Addressing) • Congestion • Flow Control • Demultiplexing @ Endpoints (Destination Address Calculation, Control/Data Sifting, Upper Layer Protocols) • Latency Variation • Security • Reliability • Heterogeneity (Hosts, Traffic Types, Nets) All Become Bigger Problems! But... 14
The Internet Community Has Solved Most of the Problems • Strengths of IP: issues of scale and heterogeneity • Weakness: Performance • ISI’s Netstation is using & promoting TCP/IP and UDP/IP • Performance problems can be solved 15
Advantages of IP • Heterogeneous Interconnects Intra-Machine Room • Wide-Area Access Enables Remote Mirroring and Backups • Future Growth Not Media-Specific • Lower R&D Investment in Networking 16
Talk Outline • Introduction • Network Technologies • NAPs in Multimedia • NAPs in Mass Storage • Operating System Support • Conclusion 17
NAPs in Multimedia Cameras, frame buffers and occasionally disk drives • ISI’s Netstation • MIT’s ViewStation • Cambridge’s Desk Area Network • HiPPI frame buffers 18
The Netstation Project Gregory Finn (project leader), Steve Hotz, Rodney Van Meter, Bruce Parham and Reza Rejaie http://www.isi.edu/netstation/ Technologies for NAPs: • Networking protocols • OS paradigms • NAP security • Multimedia & storage 19
Netstation Netstation is a system composed of network-attached peripherals (NAPs) created by replacing the system bus in a workstation with a gigabit network. HiDef User Input Internet as Backplane CPU/Memory Camera Disk • Use Internet protocols for ubiquitous device access • Based on ATOMIC 640 Mbps switched network 20
ViewStation & Desk Area Network • Principle difference: physically-defined boundary • ATM Hi-Def Display magnetic disk DAN CPU/Memory boundary ATM to LAN/WAN gateway Network camera magnetic disk 21
Third Party Transfer • Direct device-to-device transfer Hi-Def Display magnetic disk data camera control cross- bar cross- CPU/Memory bar to LAN/WAN CPU/Memory keyboard/mouse magnetic RAM Disk disk 22
Talk Outline • Introduction • Network Technologies • NAPs in Multimedia • NAPs in Mass Storage • Operating System Support • Conclusion 23
NAPs in Mass Storage • SGI Origin 2000? • CMU Network-Attached Secure Disk (NASD) • LLNL’s Network-Attached Peripheral (NAP) RAID • Fibre Channel Disk Drives • Palladio at HP Labs • Petal/Frangipani at DEC • Global File System from UMinn • National Storage Industry Consortium’s NASD Committee http://www.hpl.hp.com/SSP/NASD/ 24
Network Disk Services Should a drive present a SCSI (block) model, or NFS (file) model, or something in between? • Low-level interface easily supports other uses (non-Unix file systems, databases, swap space, network RAID) • File model may distribute functionality more widely, scaling better • Architectural tradeoffs are complex 25
CMU Network Attached Secure Disk Group • Defined useful taxonomy • Their disks hold “objects”, like unnamed NFS files • File manager/name service centralized • http://www.pdl.cs.cmu.edu/NASD/ workstation cross- bar workstation file manager data magnetic magnetic disk disk control 26
Talk Outline • Introduction • Network Technologies • NAPs in Multimedia • NAPs in Mass Storage • Operating System Support • Conclusion 27
Operating System Issues with NAPs • Resource discovery • Concurrency/sharing • Security • Programming paradigms for third-party transfer 28
Security • Access not physically constrained • Cryptographic authentication required • Who a request comes from is more important than where • Devices don’t understand “users” • Netstation approach: Derived Virtual Devices (DVDs) 29
Third-Party Transfer • read/write paradigm inadequate -- generalize to move(source,destination) • Concurrency management • Error handling: to partner, requestor or owner of one or both devices? • Details: boundary conditions, blocking factors, generalized RPC formats 30
Conclusions • Network Attached Peripherals (NAPs) allow new system architectures More scalable interconnects Direct device communication • Key issues: Security Scale Performance Legacy • “A Brief Overview of Current Work on Network Attached Peripherals”, ACM OSR Jan. ‘96 or web page below • http://www.isi.edu/~rdv/ 31
Recommend
More recommend